Evaluations of school-based drug prevention programs are criticized for using individual as the unit of analysis because the analysis does not take the intraclass correlation (ICC) into account. Critics often recommend that the school be the unit of analysis. However, using school as the unit of analysis is undesirable for two important reasons: The analysis is too conservative, and important processes may have meaning only at the individual (or class) level of analysis. Forcing the analysis to the school level may yield analyses that are devoid of substantive meaning. We recommend using an analysis that takes the hierarchical structure of the data into account: Multi level analysis (also known as hierarchical linear modeling and random coefficient analysis). This analysis, which can examine individuals within schools, or individuals within classrooms within schools, etc., controls for the ICC. The first major goal of the proposed research is to make use of this state-of-the- art analytic strategy to reassess the drug prevention effectiveness of two major prevention programs (Projects SMART and AAPT). Multi level analysis also allows the researcher to test important interactions between variables that have meaning at different levels of analysis. For example, programs that are implemented well should have better prevention outcomes than are those that are implemented poorly. This would be a three level analysis: Program (a school-level variable) interacts with quality of program implementation (a class-level variable) to affect drug use outcomes (an individual level variable). This research will also test important multilevel interactions involving prevention program effects on adolescent drug use outcomes. The proposed research will also explore two new areas of research involving multilevel analysis: Latent variable multilevel analysis, and multilevel analysis with some data missing. Latent variable multilevel analysis will be explored by making using of procedures described recently by Muthen and by McDonald (BIRAM). Missing data procedures to be explored will be (a) a multiple imputation approach (Rubin, 1987), (b) an approach (involving existing multilevel software) that treats the data has having the structure of multiple responses within individuals within schools, and (c) a new EM algorithm based on recent work by Little & Rubin (1987) and McDonald (in press).
| Palmer, R F; Graham, J W; White, E L et al. (1998) Applying multilevel analytic strategies in adolescent substance use prevention research. Prev Med 27:328-36 |
| Kreft, I G (1998) An illustration of item homogeneity scaling and multilevel analysis techniques in the evaluation of drug prevention programs. Eval Rev 22:46-77 |
